GUIDES
AI Will Train Itself by 2028 đź“š
Anthropic co-founder Jack Clark just gave us a 36-month calendar . Here’s what he actually said, why it’s the most useful timeline anyone trying to learn AI.
The Quote
On May 4, 2026, Anthropic co-founder Jack Clark — the head of Policy at the company that builds Claude — published an unusually direct prediction in his Import AI newsletter. Most coverage framed it as a doom warning. The cleaner read is that he just handed everyone trying to learn AI a precise calendar.
Jack Clark, Anthropic co-founder
“I now believe there’s a likely chance (60%+) that no-human-involved AI R&D … happens by the end of 2028.”
“Recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves.”
Source: Import AI #455 (May 4, 2026)
Translation: an Anthropic co-founder — not a doomer pundit, not a journalist — says there’s a better-than-coin-flip chance that within ~3 years, AI systems will be capable of building their own successors with no human in the loop. The threshold he’s describing is “an AI system powerful enough that it could plausibly autonomously build its own successor.”
Why He Believes It (The Data)
The 60% number isn’t a vibe. It’s anchored in measurable progress on how long an AI agent can work autonomously before something breaks — the metric researchers call “time-horizon length.” The independent eval lab METR has been tracking this for years and the curve is accelerating.
Why This Is a Gift, Not a Threat
Most major tech shifts blindside people. Nobody got a 3-year warning before the iPhone, or social media, or the cloud. Career-relevant info on what was about to change came in fragments — mostly to people already inside the industry.
This is the first time a major curve has been forecast out loud, by the people building it, with a number, a horizon, and a published timeline. That’s a gift. The people who treat the next 36 months like a calendar instead of a cliff are the ones who’ll be in position when it lands.
5 Things to Start Learning Right Now
Not predictions. Not vague advice. Five specific skills that compound regardless of which company “wins” AI in 2028.
By 2028, the basic unit of AI-mediated work is an agent — not a prompt. The 4-part structure (Context, Connections, Workflows, Memory) is going to be table stakes for everyone’s job. Build one in Claude Code this month so the structure is muscle memory before everyone else catches up.
As model output gets better, the bottleneck moves from “can it write this?” to “how do you know if it wrote it well?” Learn to write rubrics, build evals, and judge AI outputs ruthlessly. Anthropic’s AI Fluency cert teaches the framework — the Discernment competency is the one that matters most here.
By 2028, AI value will live in the systems your agents are connected to — not the model itself. Pick one stack you care about (Gmail + Calendar + Drive, or Shopify + Klaviyo + Meta, or your CRM + Slack + a database) and become the person who knows how to wire AI through it cleanly.
Tools like Lovable and Claude Code let you ship a working app in a weekend with no engineering background. The proof you can do this is the resume line that’ll matter most in 2027. Build one app, one Claude Skill, one agent — and put it where someone can click it.
Whatever the 2028 model can do will be shaped at least as much by what it’s allowed to do. The people who understand the policy frontier — Responsible Scaling Policies, model evaluations, the EU AI Act, frontier-AI safety research — will be the rare hire that bridges “can build with it” and “can run a company that uses it responsibly.”
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.